Feedforward neural networks can be used for nonlinear dynamic modeling. Although the basic principle of employing such networks is straightforward, the problem of selecting the training data set and the network topology is not a trivial task. This paper examines the use of genetic algorithm optimization techniques to optimize the neural network. The paper presents the results of studies on the effect of number of neurons and input combination method on the performance of neural networks and the application of this study to improve leak monitoring in pipelines. The neural networks examined in this study do not use the sensor reading directly as in conventional neural networks but combine it using polynomial type laws to produce hybrid inputs. The optimization technique tries to find the best polynomial laws (input combination) to reduce network size, head variation effect, and optimize network performance. The resulting networks show superior performance and use fewer numbers of neurons.

This content is only available via PDF.
You do not currently have access to this content.